MS Lesion Finder Tool

The Multiple Sclerosis (MS) Lesion Finder Tool

The MS Lesion Finder is a specialised version of the Fuzzy Connector. It has been specially tailored to the task of finding abnormalities in the (predominantly white matter) brain parenchyma. This is a task which is widely performed as part of many clinical studies of MS. The tool uses two template images that are built into Jim that incorporate detailed knowledge of the likely distribution and size of MS lesions in the head. The lesion prior probability template image has been constructed from a large sample of MS patients' scans, and is used by the MS Lesion Finder rather like a "fuzzy mask", to reduce the fuzzy affinity to pixels where it is unlikely that MS lesions are found. Using this tool should make the task of quantifying MS lesions more reliable, less labour intensive and and less operator dependent.

You can, in principle, use the MS Lesion Finder to find lesions in any type of MRI scans of the head, including proton-density-weighted, T₂-weighted, FLAIR and T₁-weighted. However, the tool has been validated using images from a double-echo sequence with the following characteristics:

Repetition time (TR): 2200-2800 ms.
Echo times (TE): 15-50 ms and 80-120 ms.
Slice thickness 3 mm.
In-plane resolution ~1mm.
Axial orientation.

First, set up the MS Lesion Finder tool as described in the introduction. If you an use an input set of images with more than one type of contrast, each image contrast must be in a separate image file. You may need to use the Slice Extractor, Image Concatenator, or Image Interleaver to achieve this. Next, set the following:

The weight for the prior probabilities of MS lesion locations, using the slider:
The default weight is 0.5. Setting a higher value will give more weight to the prior probabilities. With a setting of 1.0 (which is, of course, not recommended), the intensities of the images are completely ignored. With a setting of 0.0, the prior distribution of MS lesions is completely ignored. In practice, you should find that a setting of between 0.25 and 0.5 works well. You should use the following as guide to selecting the weight:
- If the segmented lesions spread too much into the surrounding grey matter, then the weight is too low. Set the weight higher.
- If there is clear variability in the extent of the segmented lesions, then the weight is too high. Typically, some lesions are segmented well, but other lesions of similar intensity characteristics are not segmented. Set the weight lower.
You will see below some example images showing the effect of varying the weight.
When selected, connections are considered in three dimensions (in-plane and between slices). When not selected, only in-plane connections are considered. Use this option when you want to delineate the lesions across multiple slices, but you do not want to define seed ROIs in all slices. It may be necessary to turn this option off when odd-numbered and even-numbered image slices are slightly out of register, as can happen if a patient moves slightly during the scan.
The MS Lesion Finder tool assumes that the image slices are ordered according to radiological convention (i.e., increasing slice number from inferior to superior for axial slices, from left to right for sagittal slices, and from anterior to posterior for coronal slices). If this is not the case, then select this option to tell the tool to reverse the slice order before analysing. This option is necessary because part of the analysis pipeline involves registering the MRI scans to a standard head template image which has radiological slice ordering.

Clicking the

button will save the setup so that next time the MS Lesion Finder tool is started, it will have the same setting as in this session.

The quickest way to define the seeds is to use Marker ROIs. Start the ROI Toolkit, change the display layout to display a single slice, and on each image slice place a Marker ROI on each lesion, as illustrated in the figure below. To place a Marker ROI, simply point to the location where you want the Marker and press the 'm' key on your keyboard. See the notes about keyboard shortcuts if this does not work for you.

Note: try place the Marker reasonably centrally within each lesion. If the lesion is large, consists of two or more confluent lesions, or contains areas that are less distinct, you can place as many markers as needed to ensure that the whole of the lesion is marked. However, Markers should not be placed outside the bounds of a lesion.

Note: you must have either one slice displayed, or a selected slice in order to be able to place ROIs.

Page through all the slices of the scan, placing Markers on all lesions in every slice. The more care you take over this task, the more reliable will be the results.

When you are ready to find the MS lesions, click the button. A series of dialogs will pop up to show that the MS Lesion finder is working. When the MS Lesion Finder is finished, it will create a set of ROIs that surround the MS lesions. These ROIs will either be:

Loaded into the displayed image, if you have selected . If you are not happy with the MS Lesion Finder's result, you can simply select Undo from the ROI Toolkit's Edit menu (or type CONTROL-Z to undo the addition of the resulting ROIs), change the MS Lesion Finder's settings, and try again.
OR
Saved to an ROI disk file, if you have selected . A File Chooser will pop up, so that you can select a file in which to save the resulting ROIs when you press the button.

Illustrated below is the result of using both echoes from a double-echo dataset as input images. The resulting ROIs are shown superimposed on the first echo image.

MS lesions outlined by the MS Lesion Finder. Both the first and second echoes of a double echo pulse sequence have been used when propagating from the seed points. Note that the smaller or less distinct lesions have been largely ignored by the finder. Setting a lower fuzzy threshold will result in larger, more inclusive ROIs.

Notes:

The intensity hint "Brighter" was provided for the first echo image, since the lesions are always brighter than the surrounding parenchyma and CSF.
The intensity hint was set to "Unknown" for the second echo image, since the lesions are darker than CSF, but brighter than parenchyma in a T₂-weighted image.
The threshold setting of 0.7 used was derived by trial and error. Since your image data is likely to have different contrast from that used to validate the tool, you will need to experiment to find the optimal fuzzy threshold setting for your data. One the optimal setting has been found, you should stick to that setting for all analyses where results are to be compared.
If you select , connections are made in 3-dimension and it may not be necessary to mark the lesions in every single slice. In principle, this should make the lesion finding more reliable and reproducible. In practice, however, slight misregistration between slices because of patient motion during the scan can give problems. The figure below shows the result when placing Marker seeds in just one slice (the slice in the centre), but connecting in 3-D. The lesions have been propagated to other slices by through-slice connections.

MS lesion propagation in 3-D.

The resulting ROIs can be edited or deleted, or new ROIs added as you wish. You can then use the ROI totaliser to compute the final lesion volume.

Note: the Markers that you used as seed points are just points on the image having no area, and therefore they do not contribute to the total lesion volume.

You should save the Markers and resulting ROIs to disk as a permament record of the procedure.

Varying the Weight on the Prior Probabilities

As noted above, setting a non-zero weight on the prior probabilities using the slider helps to control unwanted spread of the ROIs, particularly into the adjacent gray matter.

The sequence of segmentations obtained below was obtained with steadily increasing weight on the prior probability. With a low weight (less than 0.3), the seeds have propagated into the grey matter adjoining the lesions. However, as the weight is increased above 0.5, it becomes more and more apparent that the shape of resulting ROI does not match the perceived border of the lesion. At a high weight, the ROI takes on the shape of typical lesions in the template probability image, rather than the shape of the lesion in this patient scan.

In general, we would recommend you use a weight of between 0.25 and 0.5 as a starting point when determining the optimal settings for your data-set.

Marker seeds

Weight=0.0 Threshold=0.55

Weight=0.1 Threshold=0.50

Weight=0.2 Threshold=0.45

Weight=0.3 Threshold=0.41

Weight=0.4 Threshold=0.40

Weight=0.5 Threshold=0.38

Weight=0.7 Threshold=0.25

Weight=0.9 Threshold=0.15

Jim Home


Marker seeds	Weight=0.0 Threshold=0.55	Weight=0.1 Threshold=0.50

Weight=0.2 Threshold=0.45	Weight=0.3 Threshold=0.41	Weight=0.4 Threshold=0.40

Weight=0.5 Threshold=0.38	Weight=0.7 Threshold=0.25	Weight=0.9 Threshold=0.15